Mechanical Design of a Cartesian Manipulatorfor Warehouse Pick and Place

M. McTaggart1,2, R. Smith1,2, J. Erskine1,2, R. Grinover1,2, A. Gurman1,2, T. Hunn1,2, N. Kelly-Boxall1,2,D. Lee1,2, A. Milan1,3, D. Morrison1,2, T. Pham1,3, G. Rallos1,2, A. Razjigaev1,2, T. Rowntree1,3, K. Vijay1,3,

S. Wade-McCue1,2, A.W. Tow1,2, Z. Zhuang 1,4, J. Leitner1,2, I. Reid1,3, P. Corke1,2, and C. Lehnert2

Abstract— Robotic manipulation and grasping in clutteredand unstructured environments is a current challenge forrobotics. Enabling robots to operate in these challenging envi-ronments have direct applications from automating warehousesto harvesting fruit in agriculture. One of the main challengesassociated with these difficult robotic manipulation tasks is themotion planning and control problem for multi-DoF (Degreeof Freedom) manipulators. This paper presents the designand performance evaluation of a novel low cost Cartesianmanipulator, Cartman who took first place in the AmazonRobotics Challenge 2017. It can perform pick and place tasksof household items in a cluttered environment. The robot iscapable of linear speeds of 1 m/s and angular speeds of 1.5 rad/s,capable of sub-millimetre static accuracy and safe payloadcapacity of 2kg. Cartman can be produced for under 10 000AUD. The complete design will be open sourced and will beavailable at the time of ICRA 2018.

I. INTRODUCTIONEnabling robots to pick and place items within cluttered

and challenging environments has direct application to indus-tries such as e-commerce, logistics and even agriculture. Oneof the main challenges associated with these difficult roboticmanipulation tasks is the motion planning and control formulti-DoF (Degree of Freedom) manipulators [1]. This canbe difficult in scenarios where the environment is cluttered,dynamic and unstructured [2] requiring large amounts ofcomputational time to find a collision-free path in the con-figuration space of the manipulator. In this paper we arguethat designing a manipulator which reduces the complexitiesof solving the motion planning problem can lead to robustand reliable solutions for real-world deployment.

Robotics competitions are a great driver for developingrobotic solutions that are reliable in real world scenarios.Amazon hold an annual competition in which they invite 16teams from around the world to help solve their warehouseautomation problem. The competition requires teams todesign a robotic picking solution that can autonomously pickand place a variety of household items. This manipulatordesign forms part of our winning entry to the AmazonRobotics Challenge 2017 [3].

This research was supported by the Australian Research Council Centreof Excellence for Robotic Vision (ACRV) (project number CE140100016).The participation at the ARC was supported by Amazon Robotics LLC.Contact: matthew.mctaggart@roboticvision.org

1Authors are with the Australian Centre for Robotic Vision (ACRV).2Authors are with the Queensland University of Technology (QUT).3Authors are with the University of Adelaide.4ZZ is with the Australian National University (ANU).

Fig. 1. Cartman, a cartesian manipulator for pick and place applications.Cartman is composed of six degrees of freedom; three prismatic joints whichform the x, y and z axes and three revolute joints and a multi-modal endeffector for grasping a large range of items.

The Amazon Robotics Challenge involves two tasks, stow-ing and picking. Stowing is the task of picking items froma cluttered tote and placing them into a storage unit [4],emulating the task of storing items received from externalsuppliers. Picking is the task of taking items from thestorage unit and placing them into boxes that are to bedelivered to your door. Depending on the storage system, thetask essentially requires large vertical motion and horizontalplanar motion.

Based on previous experiences, the most common type ofmanipulators used for this challenge are fully articulated, aUR-5 and Baxter [5] [6] [7]. The main advantage of using anarticulated manipulator is the ratio of mechanical footprintto workspace. However, using a fully articulated manipulatorcomes with the challenges such as singularities [1] in theinverse kinematics which can cause the motion plan to fail orresult in unwanted large joint velocities [8]. A bigger issuefor robotic pick and place tasks is that the tip position is

defined by the item to be grasped which give fewer optionsfor the configuration of the arm to achieve a successful grasp.Collisions are likely in this situation. The advantage withCartesian robots is that they have the same shape irrespectiveof tool position.

Cartesian manipulators are comprised of three orthogonallinear actuators. If there is no requirement for a small me-chanical footprint or the structure of the environment is notcomplex, such as a mobile manipulation task, then Cartesianmanipulators are well suited. Planning times are faster due tosimpler inverse kinematics because the configuration spaceis the task space.

In this paper we propose a cost-effective design of aCartesian robot (seen in Fig. 1) for pick and place tasksalong with extensive performance measures. In particular,the contributions of this paper include:

• The cost-effective design of a 6 DoF Cartesian manip-ulator.

• Novel differential belt drive system to assist with singlemotors driving all distal motors.

• Extensive performance analysis verifying the design foruse in pick and place tasks

It features six degrees of freedom; three prismatic jointswhich form the x, y and z-axes and three revolute jointswhich form the wrist allowing the end effector to have aroll, pitch and yaw motion. The complete design will beopen sourced and will be made available online here: urlwill be added for camera ready version. A detailed systemoverview or our system of the Amazon Robotics Challengetask can be found here [3].

This paper presents a brief comparison between Cartesianand articulated manipulators (section II), design (section III)and performance evaluation (section IV) and a discussionof our low-cost Cartesian manipulator. The robot arm hasthe ability to be used in other scenarios in which a largeplanar workspace is available, such as fruit harvesting in agreenhouse [9] or factory pick and place tasks.

II. CARTESIAN VS ARTICULATED

Articulated manipulators are common for many differentapplications due to their versatility and large workspace withrespect to their mechanical footprint. The drawbacks are thatthey can have singularities and discontinuity in trajectory forcertain end effector configurations [10] [8]. There are waysof reducing the effect of singularities on motion planning butnot eliminating them as shown in the works of [11] and [12].Using a Cartesian manipulator with a wrist joint to work in acartesian work space eliminates almost all singularities andreduces discontinuities. A comparison between Cartman andother articulated robots can be seen in Fig. 2. These dis-continuities only affect the end-effectors if they try to rotatethrough them. Since Cartman works in a primarily verticaldirection without rotating the wrist joint, these discontinuitiesdo not affect regular operation of the robot. A disadvantagewith a cartesian manipulator is the requirement for a largermechanical footprint in ratio to the overall workspace of themanipulator. This is due to the fact that the linear motion

requires some form of support, usually a rail, along the entirelength of the axis.

III. DESIGN

III-A Overview

The entire manipulator system is mounted on an alu-minium stand shown in Fig. 4. It has three linear axes and awrist joint which consists of a roll, pitch and yaw axis. Alsofeatured in Fig. 4 is the gripper [13].

III-B Specifications

When designing the manipulator for the Amazon RoboticsChallenge task we set out the following specifications:

• A reachable workspace of 1.2m × 1.2m × 1.0m• A top linear velocity of 1m/s under load along the three

linear axes (x,y and z).• A top angular velocity of 1rad/s under load in the

angular axes (roll, pitch and yaw).• Have a load capacity of 2kg.• Six DoF at the end-effector, given by three linear axes

forming the Cartesian gantry and a three axis wrist.• Be able to be easily deconstructed/reconstructed for

transportation overseas to the Amazon Robotics Chal-lenge event location.

III-C Mechanical Design

Frame: The frame design is constructed out of laser cutand folded 1.2mm sheet aluminium. This was used to keepweight down for transport. The sheet aluminium formedthe main outer frame for the system and housed the x-axisbelt system and transmission rod. The manipulator frame ismounted on an aluminium stand as seen in Fig. 4.

Motors: Linear motion is performed using Technic’sClearPath SD-SK-2311S motors. They are a closed-loopbrushless motor system designed to be a drop in replacementfor stepper motors. By using a stepper-like motor, eliminatesthe need for external encoders, and makes controlling thelinear axes quite simple and similar to how a 3D printer iscontrolled. They were chosen due their high performance andease of use.

For the roll, pitch and yaw axes, three Dynamixel ProL54-50-500 are used. These motors provide the necessaryoperating torque to hold a 2kg on the end effector whilstunder acceleration.

Power Transmission: To actuate the prismatic joints, abelt and pulley system was used. A single motor drivesthe x axes. In order to eliminate a cantilever effect on thex axes, a transmission rod is used to transmit power fromone member to another. One common design that has beenobserved in a lot of simple manipulator designs, is eachaxis motor needs to carry the weight of all distal motorsas well as the payload. As a result, more powerful motorsare needed which then increases weight as well as cost. Tosolve this problem a differential belt system was designed.Rather than using a single motor to drive a single axes, two

Fig. 2. Comparison of discontinuity maps between Cartman, a Baxter and UR-5. (a) Map of Cartman’s sucker end-effector, (b) Map of Baxter’s leftgripper, (c) Map of the UR-5 end-effector. If the end-effector passes through these boundaries, joint velocities can accelerate to infinity. Cartman’s designlimits these discontinuity boundaries, making planning simpler and safer.

Fig. 3. Differential belt system implemented for Y and Z-axes. The resulting motion of the arm is the sum of the individual belt vectors. In this example,the four middle idler pulleys are able to move in the vertical plane. Diagonal motion is obtainable by combination of two of the examples.

motors work in tandem to drive two axes. The system canbe seen in Fig. 3.

Rails and Bearings: The linear rails used for the x andy axes are the TBR20 and TBR15 respectively. These area precision type profile rail which offer a higher precisioncompared to other linear type rails. Smaller rails were usedin the y axis to reduce weight on the system. The y axisconsists of two 10mm round rails. Although 10mm rails aresmall, they offer a light weight solution to the problem. Thedownside to using 10mm rail however is that when the zaxis is extended it creates a pendulum out of the system andinduces oscillations at the end effector due to deflection.This was a trade off that was considered during the designprocess. Although deflection and oscillations are present,steady state accuracy is still achieved with ease once settled.

III-D Electrical Design

A single microcontroller is used to control all six axes.The microcontroller that was chosen was the Teensy 3.6 asmany libraries exist and has a suitable processor capableof over 180MHz operation. The breakout board needed to

include a logic shifter circuits for each of the ClearPathmotor pins as the input and output threshold for thismicrocontroller is only 3.3V and the ClearPath motorsneeded minimum 5V pulse to operate. In order to interfacewith the Dynamixel Pro motors, an RS485 module was usedso that the Teensy could communicate using the DynamxielPro protocol.

III-E Software Design

The software can be broken down into two sections:software used on the PC and the firmware used on board themanipulator. A system block diagram of different softwareinteraction can be seen in Fig. 5. Robot Operating System(ROS) [14] was used to handle the high level functionalityof the system. Desired end effector poses and robot stateswere published using the ROS MoveIt! package [15]. Thiswas published as a ROS JointState message type which wasprocessed by the Teensy.

The low level firmware functions send commands to boththe ClearPath and Dynamixel Pro motors and also readany feedback that was available from the motors. As theClearPath motors are a drop in replacement for regular

Fig. 4. Isometric view of the entire manipulator. Key components have beenlabelled and are as follows: (A) Aluminium T-slot stand, (B) ManipulatorAluminium frame, (C) X-Axis TBR20 profile rails, (D) Y-Axis TBR15profile rails, (E) Z-Axis 10mm round rails, (F) YZ motor-carriage, (G)Suction gripper, (H) Wrist, (I) Parallel plate gripper

Fig. 5. System block diagram of the communication between software andfirmware packages of the manipulator system.

stepper motors, a stepper motor library was able to beused. The library which was used is AccelStepper [16],a library for use with Arduino compatible devices. Thisexpands on the standard Stepper library provided by Arduino.The main benefit in using AccelStepper is that it allows theuser to control the acceleration profile of the motor whichis desirable for this application. The Dynamixel Pro werecontrolled using a modified version of their OpenCR library[17]. The public open source library was developed to beused with their own electronic hardware. Slight modificationswere made in order to adapt the library for use with theTeensy 3.6.

IV. PERFORMANCE EVALUATION

IV-A Methodology

In order to test whether the design specs have been met,the motion of the system was tested. Motive, a motioncapture system was used to track trajectories during differentpath plans. The setup can be seen in Fig. 6. Each axiswas tested individually at various heights and speeds to

Fig. 6. Motion capture setup. Five motion capture cameras (red boxes)were used to determine the pose of the multi-reflective body (blue box)

Fig. 7. Error in position of the x-axis. The x-axis is driven at fast and slowspeed at a high and low z-height. (a) high-fast, (b) high-slow, (c) low-fast,(d) low-slow.

test the end point accuracy during motion. Each axis wascommanded at a fast and slow speed moving to and fromthe extremes of the tracking system. The pose of the multi-reflective body seen in Fig. 6 is tracked and interpolated tocompare the desired joint position with recorded positionwhilst in motion and for steady state error.

IV-B Results

A summary of the tests conducted can be found inTable I. Below are the plots of that data that was capturedand summarised in these tables.

X-axis: The results of X-axis movements can be seen inFig. 7. The steady can be seen to converge to close to zeroonce motion has ceased. The average error in position is9.62 mm as seen in I. During slow movements, oscillationsoccur about the desired point due to the pendulum effect butstill converge to the desired end point. This is consideredsufficient for the application that the manipulator is designedfor.

Y-axis: As seen in Fig. 9, the y-axis trajectory oscillatessubstantially at low heights at slow speeds. This is causedby the pendulum effect of the system as described in section

Fig. 8. Wrist trajectory at fast and slow speeds. (a) roll-fast, (b) roll-slow, (c) pitch-fast, (d) pitch-slow, (e) yaw-fast, (f) yaw-slow.

TABLE ISUMMARY OF ERROR ACROSS SINGLE AND MULTIPLE AXES

StandardDeviation

Mean Error Static Error

x 16.07 9.62 6.2e−01 mmy 13.73 9.42 1.9e−01 mmz 13.70 7.91 1.9e−03 mmxzy 64.88 29.63 1.5e−01 mmroll 0.02 0.01 5.6e−04 radpitch 0.02 0.01 1.5e−04 radyaw 0.02 0.01 7.3e−05 rad

Each axis is driven individually and simultaneous motion of axes arealso evaluated (xyz). Mean error is defined as the error between thedesired position and the measured position. Static error is the errorbetween desired position and measured position once motion has ceased.

Fig. 9. Error in position of the y-axis. The y-axis is driven at fast and slowspeed at a high and low z-height. (a) high-fast, (b) high-slow, (c) low-fast,(d) low-slow.

III. For the task of the Amazon Robotics Challenge, finalpoint accuracy is more crucial than in-motion accuracy. Wecan see that the final desired position converges to almostzero. Given this specific task, slow y-axis movements are notnecessary any way. The coupling effect of the differentialbelt system is also of interest. The trajectory of the z-axiswhile moving the y-axis can be seen in Fig. 10. We cansee that the z-axis does travel during pure y-axis motionas a result of the differential belt system. This is to beexpected as the two ClearPath motors need to be perfectlyin sync in order to reduce this movement to zero. Despitethis travel, the final point accuracy is still sufficient for themanipulator’s intended task.

Z-axis: The z-axis can follow a pure Z trajectory witha fair amount of accuracy as seen in Fig. 11. The z-axis

Fig. 10. Z-axis trajectory during pure y-axis motion. The oscillations area result of the pendulum effect. (a) high-fast, (b) high-slow, (c) low-fast,(d) low-slow.

Fig. 11. Z-axis trajectory at various speeds. (a) high-fast, (b) high-slow,(c) low-fast, (d) low-slow.

exhibits the least amount of error while in motion with only7.9 mm mean error in position. As discussed previously,due to the coupling of the Y and Z axes, the y-axis wandersduring pure z-axis motion just as the z-axis wanders duringpure y-axis motion. Despite this, steady state precision isstill achieved.

Wrist - roll, pitch and yaw: The wrist joint of the robotexhibit exceptional accuracy not only in motion but insteady state as well as seen in Fig. 8. Achieving a steadystate of almost zero error.

XYZ axes: Simultaneous motion of multiple axes weretested against a more complex trajectory. A lemniscate curvewas generated on firstly a single plane at three differentheights and also in a multi-plane test. The system was testedat different speeds and exhibited the same characteristicsas the single axes. Steady state error is almost zero seen inFig. 12.

Fig. 12. Three axis movements in the X, Y, and Z axes. (a) xyz-fast, (b)xyz-slow.

Fig. 13. Example trajectory of a real pick. The task was to pick up a 1kgdumbbell from a start point, place it in another area and return to the startposition

Example Pick run: To simulate a pick task within theAmazon Robotics Challenge a simple scenario was set upin order to test accuracy during motion the manipulatorwas intended for. The task was to pick an item from theAmazon item set, in this case a 2lb weight, and place it toanother location and then return to its start position. Thetrajectory recorded can be seen in Fig. 13. The manipulatorwas easily able to achieve the task with a mean error ofonly 2.96mm during motion.

IV-C Verification

Using the data presented in this section, we can verifyour speed specifications. Unfortunately due to mechanicallimitations, the system experienced belt slip under highaccelerations. As a result the acceleration parameter ofthe controller needed to be lowered. Table II provides asummary of the speed and error achieved by each axis. We’veshown that the system is capable of moving a load whilstfollowing linear trajectories with reasonable accuracy. TableI features the static error for each axis. We found that despiteoscillations and trajectory error during motion, the steadystate error is negligible and can achieve sub-mm and sub-mrad accuracy.

V. DISCUSSION AND CONCLUSION

As seen in section IV-B we’ve shown the limitations ofthe system in terms of trajectory motion. Y axis motioncontributes the most to the overall error due to the pendulumeffect of the system. We’ve shown that this only happens atlow speeds at approximately speeds of 0.1m/s or less. Toavoid this behaviour, this limits the operating speeds forcingfast movements possibly limiting the applications of whichthis design can be used. The maximum target speed was

TABLE IISUMMARY OF SPEEDS ACHIEVED BY THE SYSTEM

Axis Commanded Speed Achieved Speed % Error

x 0.604 0.572 5%y 0.531 0.493 7%z 0.512 0.472 8%roll 1.547 1.687 -9%pitch 1.536 1.531 0%yaw fast 1.593 1.660 -4%

As the controller used for the system is a position controller, negativeerror values indicate that the joint is able to track cope with latency andaccurately track a trajectory.

not able to be reliably achieved due to belt slippage. As aresult, the acceleration of the system was limited yieldinga maximum achieved speed of only 0.57 m/s. In future wewish to address this issue as well as the slow movementoscillations by increasing the z-axis rail and bearing system.

REFERENCES

[1] L. Sciavicco and B. Siciliano, Modelling and control of robot manip-ulators. Springer Science & Business Media, 2012.

[2] A. H. Qureshi, K. F. Iqbal, S. M. Qamar, F. Islam,Y. Ayaz, and N. Muhammad, “Potential guided directional-RRT* for accelerated motion planning in cluttered environments,”in 2013 IEEE International Conference on Mechatronics andAutomation. IEEE, aug 2013, pp. 519–524. [Online]. Available:http://ieeexplore.ieee.org/document/6617971/

[3] D. Morrison, A. W. Tow, et al., “Cartman: The low-cost cartesianmanipulator that won the amazon robotics challenge,” AustralianCentre for Robotic Vision, Tech. Rep., 2017.

[4] Amazon Robotics, “AR Challenge :: Ama-zon Robotics,” 2017. [Online]. Available:https://www.amazonrobotics.com/#/roboticschallenge/rules

[5] “UR5 - The flexible and collaborative robotic arm.”[6] “Baxter Collaborative Robots for Industrial Automation — Rethink

Robotics.” [Online]. Available: http://www.rethinkrobotics.com/baxter/[7] N. Correll, K. E. Bekris, D. Berenson, O. Brock, A. Causo, K. Hauser,

K. Okada, A. Rodriguez, J. M. Romano, and P. R. Wurman, “Lessonsfrom the Amazon Picking Challenge,” jan 2016. [Online]. Available:http://arxiv.org/abs/1601.05484

[8] P. I. Corke, Robotics, vision and control : fundamental algorithms inMATLAB. Berlin, Heidelberg: Springer, 2011.

[9] C. Lehnert, A. English, C. Mccool, A. W. Tow, and T. Perez, “Au-tonomous Sweet Pepper Harvesting for Protected Cropping Systems,”IEEE Robotics and Automation Letters, vol. 2, no. 2, pp. 872–879,2017.

[10] C. Innocenti and V. Parenti-Castelli, “Singularity-free evolution fromone configuration to another in serial and fully-parallel manipulators,”Journal of Mechanical Design, vol. 120, no. 1, pp. 73–79, mar 1998.

[11] K. Hauser and Kris, “Continuous Pseudoinversion of a MultivariateFunction: Application to Global Redundancy Resolution,” Workshopon the Algorithmic Foundations of Robotics, jan 2017. [Online].Available: https://dukespace.lib.duke.edu/dspace/handle/10161/13487

[12] Y. Nakamura and H. Hanafusa, “Inverse Kinematic Solutions WithSingularity Robustness for Robot Manipulator Control,” Journal ofDynamic Systems, Measurement, and Control, vol. 108, 1986.

[13] S. Wade-McCue, N. Kelly-Boxall, et al., “Tbc — design of a multi-modal end-effector and grasping system how integrated design helpedwin the amazon robotics challenge,” Australian Centre for RoboticVision, Tech. Rep. ACRV-TR-2017-03, 2017.

[14] M. Quigley, B. Gerkey, K. Conley, J. Faust, T. Foote, J. Leibs,E. Berger, R. Wheeler, and A. Ng, “Ros: an open-source robotoperating system,” in Proc. of the IEEE Intl. Conf. on Robotics andAutomation (ICRA) Workshop on Open Source Robotics, Kobe, Japan,May 2009.

[15] I. A. Sucan and S. Chitta, “MoveIt!” [Online]. Available:http://moveit.ros.org/

[16] M. McCauley, “AccelStepper: AccelStepper li-brary for Arduino,” 2010. [Online]. Available:http://www.airspayce.com/mikem/arduino/AccelStepper/

[17] R. L. W. Jung, “Robotis OpenCR.”[Online]. Available: https://github.com/ROBOTIS-GIT/OpenCR/tree/master/arduino/opencr arduino/opencr/libraries/DynamixelSDK